Communication methods and devices, and storage medium and program product

Abstract

The present disclosure relates to communication methods and devices, and a storage medium and a program product. A method, which may be executed by a terminal, comprises: sending first information to a network device, wherein the first information is used for indicating a first model structure, and the first model structure is a model structure for which the network device is required by a terminal to transmit a model parameter; and receiving second information sent by the network device, wherein the second information is used for indicating the model parameter of a first model, and the model structure of the first model is the first model structure. Therefore, the resource overheads of model parameter transmission can be effectively reduced, thereby effectively improving the system performance.

Description

Communication methods, devices, storage media and software products Technical Field

[0001] This disclosure relates to the field of communication technology, and in particular to communication methods, devices, storage media and program products. Background Technology

[0002] With the development of artificial intelligence (AI) and machine learning (ML) technologies, AI technology can reduce the feedback overhead of terminals or improve the accuracy of CSI feedback. For example, a bilateral AI / ML model can be used to generate a partial model based on the terminal-side channel state information (CSI) and a partial model based on the network-side CSI recovery model to achieve compressed feedback and recovery of CSI respectively.

[0003] Summary of the Invention

[0004] This disclosure provides a communication method, device, storage medium, and program product.

[0005] According to a first aspect of the embodiments of this disclosure, a communication method is provided, executed by a terminal, the method comprising:

[0006] Send first information to the network device, the first information being used to indicate a first model structure, the first model structure being a model structure for which the terminal needs the network device to transmit model parameters;

[0007] The system receives second information sent by a network device, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0008] According to a second aspect of the embodiments of this disclosure, a communication method is provided, performed by a network device, the method comprising:

[0009] The receiving terminal sends first information, which is used to indicate a first model structure, and the first model structure is a model structure for which the terminal needs the network device to transmit model parameters.

[0010] Send a second message to the terminal, the second message being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0011] According to a third aspect of the embodiments of this disclosure, a communication device is provided, comprising:

[0012] The transceiver module is used to send first information to the network device. The first information is used to indicate a first model structure. The first model structure is a model structure for which the terminal needs the network device to transmit model parameters.

[0013] The transceiver module is further configured to receive second information sent by the network device, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0014] According to a fourth aspect of the embodiments of this disclosure, a communication device is provided, comprising:

[0015] The transceiver module is used to receive first information sent by the terminal. The first information is used to indicate a first model structure, which is a model structure for which the terminal needs the network device to transmit model parameters.

[0016] The transceiver module is further configured to send second information to the terminal, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0017] According to a fifth aspect of the embodiments of this disclosure, a communication device is provided, comprising:

[0018] One or more processors;

[0019] The communication device is used to perform the communication method described in the first or second aspect.

[0020] According to a sixth aspect of the present disclosure, a communication system is provided, including a network device and a terminal, wherein the terminal is configured to implement the communication method described in the first aspect, and the network device is configured to implement the communication method described in the second aspect.

[0021] According to a seventh aspect of the present disclosure, a storage medium is provided that stores instructions which, when executed on a communication device, cause the communication device to perform the communication method as described in the first or second aspect.

[0022] According to an eighth aspect of the present disclosure, a computer program product is provided, comprising a computer program and / or instructions that, when executed by a communication device, implement the communication method as described in the first or second aspect.

[0023] In the above embodiments, by sending first information to the network device through the terminal to indicate that it needs the network device to transmit the model structure of the model parameters, the network device can transmit only the model parameters required by the terminal, which can effectively reduce the resource overhead of model parameter transmission and effectively improve system performance. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings required for the description of the embodiments are introduced below. The following drawings are only some embodiments of this disclosure and do not impose specific limitations on the protection scope of this disclosure.

[0025] Figure 1A is an exemplary schematic diagram of the architecture of a communication system provided according to an embodiment of the present disclosure.

[0026] Figure 1B is a schematic diagram illustrating a CSI compression feedback and recovery based on a bilateral AI / ML model according to an embodiment of the present disclosure.

[0027] Figure 2 is an exemplary interactive schematic diagram of a communication method provided according to an embodiment of the present disclosure.

[0028] Figure 3A is an exemplary flowchart of a communication method provided according to an embodiment of the present disclosure.

[0029] Figure 3B is an exemplary flowchart of a communication method provided according to an embodiment of the present disclosure.

[0030] Figure 4A is an exemplary flowchart of a communication method provided according to an embodiment of the present disclosure.

[0031] Figure 4B is an exemplary flowchart of a model structure provided according to an embodiment of the present disclosure.

[0032] Figure 5A is an exemplary structural diagram of a terminal provided according to an embodiment of the present disclosure.

[0033] Figure 5B is an exemplary structural diagram of a network device provided according to an embodiment of the present disclosure.

[0034] Figure 6A is an exemplary structural diagram of a communication device provided according to an embodiment of the present disclosure.

[0035] Figure 6B is an exemplary structural diagram of a communication device provided according to an embodiment of the present disclosure. Detailed Implementation

[0036] This disclosure provides a communication method, device, storage medium, and program product.

[0037] In a first aspect, embodiments of this disclosure provide a communication method executed by a terminal, the method comprising:

[0038] Send first information to the network device, the first information being used to indicate a first model structure, the first model structure being a model structure for which the terminal needs the network device to transmit model parameters;

[0039] The system receives second information sent by a network device, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0040] In the above embodiments, by sending first information to the network device through the terminal to indicate that it needs the network device to transmit the model structure of the model parameters, the network device can transmit only the model parameters required by the terminal, which can effectively reduce the resource overhead of model parameter transmission and effectively improve system performance.

[0041] In conjunction with some embodiments of the first aspect, in some embodiments, the model parameters of the first model include only the model parameters corresponding to the first sub-model structure, and the first sub-model structure includes at least one of the following: a portion of the neural network layer in the first model, a portion of the neurons in the first model, a portion of the channels in the first model, and a portion of the convolutional kernels in the first model.

[0042] In the above embodiments, the network device only transmits the model parameters corresponding to the sub-model structures that are not predefined, without transmitting all the parameters of the model, which effectively reduces resource overhead.

[0043] In conjunction with some embodiments of the first aspect, in some embodiments, the model parameters of the first model are pruned model parameters.

[0044] In the above embodiments, the model can be pruned before transmitting the pruned model parameters, which can reduce the resource overhead of the model parameters.

[0045] In conjunction with some embodiments of the first aspect, in some embodiments, the second information is also used to indicate pruning content;

[0046] The pruning includes at least one of the following:

[0047] The model parameters of the first model are subject to weight pruning. The pruning includes a first neuron and / or a first neural network layer. The first neuron is the neuron whose model parameters are pruned in the first model, and the first neural network layer is the neural network layer whose model parameters are pruned in the first model.

[0048] The model parameters of the first model are pruned by neurons. The pruning includes the first neuron, which is the neuron in the first model whose corresponding model parameters are pruned.

[0049] The model parameters of the first model are pruned by convolution kernels. The pruning includes the first convolution kernel, which is the convolution kernel in the first model whose corresponding model parameters are pruned.

[0050] The model parameters of the first model are structurally pruned. The pruning includes a first neural network layer and / or a second model parameter. The first neural network layer is the neural network layer in the first model whose corresponding model parameters are pruned. The second model parameter is the model parameter between the neural network layers in the first model that is pruned.

[0051] The model parameters of the first model are pruned through channels. The pruning includes the first channel, which is the channel in the first model where the corresponding model parameters are pruned.

[0052] In the above embodiments, the network device can indicate the pruning content corresponding to different pruning types through the second information, which enables the terminal to configure the model parameters more accurately and ensures the reliability of the model configuration.

[0053] In conjunction with some embodiments of the first aspect, in some embodiments, the model parameters of the first model are quantized model parameters;

[0054] In this model, the neural network layer or neurons are divided into multiple parts, and the model parameters corresponding to each part correspond to different quantization precisions.

[0055] In the above embodiments, model parameters can be quantized, and the neural network layers or neurons of the model can be divided into multiple parts, with different quantization scales used for different parts. This effectively reduces the resource overhead of parameter transmission while ensuring the accuracy of parameter transmission.

[0056] In conjunction with some embodiments of the first aspect, in some embodiments, the first model includes multiple models with different model structures, the multiple models include a second sub-model structure with the same model structure, the model parameters corresponding to the second sub-model structure are the same, and the second sub-model structure includes at least one of the following: neural network layer; neuron; channel; convolution kernel.

[0057] In the above embodiments, multiple models can be designed to include the same sub-model structure, which eliminates the need to transmit all model parameters of each model and effectively reduces resource overhead.

[0058] In conjunction with some embodiments of the first aspect, in some embodiments, the second information is further used to indicate the structure of the second sub-model; or,

[0059] The terminal negotiates with the network device to determine the structure of the second sub-model; or...

[0060] The terminal reports the structure of the second sub-model.

[0061] In the above embodiments, network devices and terminals can reach a consensus on which sub-model structures to use for multiple models through network instructions, terminal reporting, or network-terminal negotiation, thereby ensuring the accuracy of parameter transmission.

[0062] In conjunction with some embodiments of the first aspect, in some embodiments, the transmission rank is greater than 1, and the model parameters of the first model satisfy at least one of the following:

[0063] The model parameters of the models corresponding to multiple or all data transmission layers are the same;

[0064] The models corresponding to multiple or all data transmission layers include a second sub-model structure with the same model structure and the same model parameters for the second sub-model structure.

[0065] The second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

[0066] In the above embodiments, when the transmission rank is greater than 1, the models corresponding to multiple transmission layers can be designed to use the same model parameters, or the multiple transmission layers can be designed to include the same sub-model structure. In this way, the network device does not need to transmit all the model parameters corresponding to each transmission layer, which effectively reduces resource overhead.

[0067] In conjunction with some embodiments of the first aspect, in some embodiments, the second information is further used to indicate a data transmission layer with the same model parameters, and / or, a second sub-model structure; or,

[0068] The terminal and the network device negotiate and determine a data transmission layer with the same model parameters, and / or, the second sub-model structure; or...

[0069] The terminal reports the same data transmission layer with the same model parameters, and / or the second sub-model structure.

[0070] In the above embodiments, network devices and terminals can reach a consensus on which transport layers use the same model parameters, or which sub-model structures are used in multiple transport layers, through network indication, terminal reporting, or network-terminal negotiation, thereby ensuring the accuracy of parameter transmission.

[0071] In conjunction with some embodiments of the first aspect, in some embodiments, the first information includes an identifier corresponding to each of the first model structures, the identifier being any one of the following: model identifier, encoder identifier, decoder identifier, or pairing identifier.

[0072] In conjunction with some embodiments of the first aspect, in some embodiments, the model parameters of the first model include at least one model parameter corresponding to the structure of the first model, and the second information is also used to indicate the model structure corresponding to each of the at least one model parameter.

[0073] In conjunction with some embodiments of the first aspect, in some embodiments, the second information is also used to indicate a second model structure, which is a model structure in the first model structure for which the second information does not indicate the corresponding model parameters.

[0074] Secondly, embodiments of this disclosure provide a communication method executed by a network device, the method comprising:

[0075] The receiving terminal sends first information, which is used to indicate a first model structure, and the first model structure is a model structure for which the terminal needs the network device to transmit model parameters.

[0076] Send a second message to the terminal, the second message being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0077] In conjunction with some embodiments of the second aspect, in some embodiments, the model parameters of the first model only include the model parameters corresponding to the first sub-model structure, and the first sub-model structure includes at least one of the following: a portion of the neural network layers in the first model, a portion of the neurons in the first model, a portion of the channels in the first model, and a portion of the convolutional kernels in the first model.

[0078] In conjunction with some embodiments of the second aspect, in some embodiments, the model parameters of the first model are pruned model parameters.

[0079] In conjunction with some embodiments of the second aspect, in some embodiments, the second information is also used to indicate pruning content;

[0080] The pruning includes at least one of the following:

[0081] The model parameters of the first model are subject to weight pruning. The pruning includes a first neuron and / or a first neural network layer. The first neuron is the neuron whose model parameters are pruned in the first model, and the first neural network layer is the neural network layer whose model parameters are pruned in the first model.

[0082] The model parameters of the first model are pruned by neurons. The pruning includes the first neuron, which is the neuron in the first model whose corresponding model parameters are pruned.

[0083] The model parameters of the first model are pruned by convolution kernels. The pruning includes the first convolution kernel, which is the convolution kernel in the first model whose corresponding model parameters are pruned.

[0084] The model parameters of the first model are structurally pruned. The pruning includes a first neural network layer and / or a second model parameter. The first neural network layer is the neural network layer in the first model whose corresponding model parameters are pruned. The second model parameter is the model parameter between the neural network layers in the first model that is pruned.

[0085] The model parameters of the first model are pruned through channels. The pruning includes the first channel, which is the channel in the first model where the corresponding model parameters are pruned.

[0086] In conjunction with some embodiments of the second aspect, in some embodiments, the model parameters of the first model are quantized model parameters;

[0087] In this model, the neural network layer or neurons are divided into multiple parts, and the model parameters corresponding to each part correspond to different quantization precisions.

[0088] In conjunction with some embodiments of the second aspect, in some embodiments, the first model includes multiple models with different model structures, the multiple models include a second sub-model structure with the same model structure, the model parameters corresponding to the second sub-model structure are the same, and the second sub-model structure includes at least one of the following: neural network layer; neuron; channel; convolution kernel.

[0089] In conjunction with some embodiments of the second aspect, in some embodiments, the second information is further used to indicate the structure of the second sub-model; or,

[0090] The terminal negotiates with the network device to determine the structure of the second sub-model; or...

[0091] The terminal reports the structure of the second sub-model.

[0092] In conjunction with some embodiments of the second aspect, in some embodiments, the transmission rank is greater than 1, and the model parameters of the first model satisfy at least one of the following:

[0093] The model parameters of the models corresponding to multiple or all data transmission layers are the same;

[0094] The models corresponding to multiple or all data transmission layers include a second sub-model structure with the same model structure and the same model parameters for the second sub-model structure.

[0095] The second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

[0096] In conjunction with some embodiments of the second aspect, in some embodiments, the second information is further used to indicate a data transmission layer with the same model parameters, and / or, a second sub-model structure; or,

[0097] The terminal and the network device negotiate and determine a data transmission layer with the same model parameters, and / or, the second sub-model structure; or...

[0098] The terminal reports the same data transmission layer with the same model parameters, and / or the second sub-model structure.

[0099] In conjunction with some embodiments of the second aspect, in some embodiments, the first information includes an identifier corresponding to each of the first model structures, the identifier being any one of the following: model identifier, encoder identifier, decoder identifier, or pairing identifier.

[0100] In conjunction with some embodiments of the second aspect, in some embodiments, the model parameters of the first model include at least one model parameter corresponding to the structure of the first model, and the second information is also used to indicate the model structure corresponding to each of the at least one model parameter.

[0101] In conjunction with some embodiments of the second aspect, in some embodiments, the second information is also used to indicate a second model structure, which is a model structure in the first model structure for which the second information does not indicate the corresponding model parameters.

[0102] Thirdly, embodiments of this disclosure provide a communication device, including:

[0103] The transceiver module is used to send first information to the network device. The first information is used to indicate a first model structure. The first model structure is a model structure for which the terminal needs the network device to transmit model parameters.

[0104] The transceiver module is further configured to receive second information sent by the network device, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0105] Fourthly, embodiments of this disclosure provide a communication device, comprising:

[0106] The transceiver module is used to receive first information sent by the terminal. The first information is used to indicate a first model structure, which is a model structure for which the terminal needs the network device to transmit model parameters.

[0107] The transceiver module is further configured to send second information to the terminal, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0108] Fifthly, embodiments of this disclosure provide a communication device, comprising:

[0109] One or more processors;

[0110] The communication device is used to perform the communication method described in the first or second aspect.

[0111] In a sixth aspect, embodiments of this disclosure provide a communication system comprising: a terminal and a network device; wherein the terminal is configured to perform the method described in the optional implementation of the first aspect, and the network device is configured to perform the method described in the optional implementation of the second aspect.

[0112] In a seventh aspect, embodiments of this disclosure provide a storage medium storing instructions that, when executed on a communication device, cause the communication device to perform the method as described in the optional implementations of the first and second aspects.

[0113] Eighthly, embodiments of this disclosure provide a computer program product, including a computer program and / or instructions, which, when executed by a communication device, cause the communication device to perform the method as described in the optional implementations of the first and second aspects.

[0114] In a ninth aspect, embodiments of this disclosure provide a computer program that, when run on a computer, causes the computer to perform the methods described in alternative implementations of the first and second aspects.

[0115] In a tenth aspect, embodiments of this disclosure provide a chip or chip system. The chip or chip system includes processing circuitry configured to perform the methods described according to optional implementations of the first and second aspects above.

[0116] It is understood that the aforementioned terminals, network devices, communication systems, storage media, program products, computer programs, chips, or chip systems are all used to execute the methods proposed in the embodiments of this disclosure. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0117] This disclosure provides communication methods, devices, storage media, and program products. In some embodiments, the terms "communication method" and "information processing method," "capability reporting method," etc., can be used interchangeably; the terms "communication device" and "information processing device," "capability reporting device," etc., can be used interchangeably; and the terms "information processing system" and "communication system," etc., can be used interchangeably.

[0118] This disclosure is not exhaustive, but merely illustrative of some embodiments, and is not intended to limit the scope of protection of this disclosure. Unless otherwise specified, each step in a particular embodiment can be implemented as an independent embodiment, and the steps can be arbitrarily combined. For example, a solution after removing some steps in a particular embodiment can also be implemented as an independent embodiment, and the order of the steps in a particular embodiment can be arbitrarily interchanged. Furthermore, the optional implementation methods in a particular embodiment can be arbitrarily combined; moreover, the embodiments can be arbitrarily combined, for example, some or all steps of different embodiments can be arbitrarily combined, and a particular embodiment can be arbitrarily combined with the optional implementation methods of other embodiments.

[0119] In each of the disclosed embodiments, unless otherwise specified or in case of logical conflict, the terminology and / or descriptions of the embodiments are consistent and can be referenced by each other. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0120] The terminology used in the embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure.

[0121] In this embodiment of the disclosure, unless otherwise stated, elements expressed in the singular form, such as "a," "an," "the," "the," "the," "the," "the," "the," "this," etc., can mean "one and only one," or "one or more," "at least one," etc. For example, when using articles such as "a," "an," "the," etc. in translation, the noun following the article can be understood as either a singular expression or a plural expression.

[0122] In the embodiments disclosed herein, "multiple" refers to two or more.

[0123] In some embodiments, the terms “at least one of”, “one or more”, “a plurality of”, “multiple”, etc., may be used interchangeably.

[0124] In some embodiments, the notation "at least one of A and B", "A and / or B", "A in one case, B in another", "in response to one case A, in response to another case B", etc., may include the following technical solutions depending on the situation: in some embodiments, A (execute A regardless of B); in some embodiments, B (execute B regardless of A); in some embodiments, execution is selected from A and B (A and B are selectively executed); in some embodiments, both A and B are executed. The same applies when there are more branches such as A, B, C, etc.

[0125] In some embodiments, the notation "A or B" may include the following technical solutions, depending on the situation: in some embodiments, A (execution of A regardless of B); in some embodiments, B (execution of B regardless of A); in some embodiments, execution is selected from A and B (A and B are selectively executed). The same applies when there are more branches such as A, B, C, etc.

[0126] The prefixes "first," "second," etc., used in the embodiments of this disclosure are merely for distinguishing different descriptive objects and do not impose restrictions on the position, order, priority, quantity, or content of the descriptive objects. The description of the descriptive objects is found in the claims or the context of the embodiments, and the use of prefixes should not constitute unnecessary restrictions. For example, if the descriptive object is a "field," the ordinal numbers preceding "field" in "first field" and "second field" do not restrict the position or order of the "fields." "First" and "second" do not restrict whether the "fields" they modify are in the same message, nor do they restrict the order of "first field" and "second field." Similarly, if the descriptive object is a "level," the ordinal numbers preceding "level" in "first level" and "second level" do not restrict the priority between "levels." Furthermore, the number of descriptive objects is not limited by ordinal numbers and can be one or more. For example, in "first device," the number of "devices" can be one or more. Furthermore, the objects modified by different prefixes can be the same or different. For example, if the object being described is "device", then "first device" and "second device" can be the same device or different devices, and their types can be the same or different. Similarly, if the object being described is "information", then "first information" and "second information" can be the same information or different information, and their content can be the same or different.

[0127] In some embodiments, “including A,” “containing A,” “for indicating A,” and “carrying A” can be interpreted as directly carrying A or indirectly indicating A.

[0128] In some embodiments, terms such as "time / frequency" and "time-frequency domain" refer to the time domain and / or frequency domain.

[0129] In some embodiments, the terms “in response to…”, “in response to determining…”, “in the case of…”, “when…”, “if…”, “if…”, etc., can be used interchangeably.

[0130] In some embodiments, the terms “greater than,” “greater than or equal to,” “not less than,” “more than,” “more than or equal to,” “not less than,” “higher than,” “higher than or equal to,” “not lower than,” and “above” can be used interchangeably, as can the terms “less than,” “less than or equal to,” “not greater than,” “less than,” “less than or equal to,” “not more than,” “lower than,” “lower than or equal to,” “not higher than,” and “below”.

[0131] In some embodiments, devices, etc., can be interpreted as physical or virtual, and their names are not limited to the names recorded in the embodiments. Terms such as “device”, “equipment”, “circuit”, “network element”, “node”, “function”, “unit”, “section”, “system”, “network”, “chip”, “chip system”, “entity”, and “subject” can be used interchangeably.

[0132] In some embodiments, "network" can be interpreted as devices included in a network (e.g., access network devices, core network devices, etc.).

[0133] In some embodiments, the terms "access network device (AN device)," "radio access network device (RAN device)," "base station (BS)," "radio base station," "fixed station," "node," "access point," "transmission point (TP)," "reception point (RP)," "transmission / reception point (TRP)," "panel," "antenna panel," "antenna array," "cell," "macro cell," "small cell," "femto cell," "pico cell," "sector," "cell group," "serving cell," "carrier," "component carrier," and "bandwidth part (BWP)" can be used interchangeably.

[0134] In some embodiments, the terms "terminal", "terminal device", "user equipment (UE)", "user terminal", "mobile station (MS)", "mobile terminal (MT)", "subscriber station", "mobile unit", "subscriber unit", "wireless unit", "remote unit", "mobile device", "wireless device", "wireless communication device", "remote device", "mobile subscriber station", "access terminal", "mobile terminal", "wireless terminal", "remote terminal", "handset", "user agent", "mobile client", and "client" can be used interchangeably.

[0135] In some embodiments, access network devices, core network devices, or network devices can be replaced by terminals. For example, embodiments of this disclosure can also be applied to structures where communication between access network devices, core network devices, or network devices and terminals is replaced by communication between multiple terminals (e.g., device-to-device (D2D), vehicle-to-everything (V2X), etc.). In this case, the structure can also be configured such that the terminal has all or part of the functions of the access network device. Furthermore, terms such as "uplink" and "downlink" can be replaced with terms corresponding to communication between terminals (e.g., "sidelink"). For example, uplink channel, downlink channel, etc., can be replaced with sidelink channel, and uplink link, downlink, etc., can be replaced with sidelink link.

[0136] In some embodiments, the terminal may be replaced by an access network device, a core network device, or a network device. In this case, the access network device, core network device, or network device may also be configured to have all or some of the functions of the terminal.

[0137] In some embodiments, the acquisition of data, information, etc., may comply with the laws and regulations of the country where the location is situated.

[0138] In some embodiments, data, information, etc., may be obtained with the user's consent.

[0139] Furthermore, each element, each row, or each column in the table of this disclosure can be implemented as an independent embodiment, and any combination of any element, any row, or any column can also be implemented as an independent embodiment.

[0140] Figure 1A is a schematic diagram of the architecture of a communication system according to an embodiment of the present disclosure. As shown in Figure 1A, the communication system 100 includes a terminal 101 and a network device 102. Optionally, the network device 102 may include at least one of an access network device and a core network device.

[0141] In some embodiments, terminal 101 includes, but is not limited to, at least one of the following: mobile phone, wearable device, Internet of Things device, car with communication function, smart car, tablet computer, computer with wireless transceiver function, virtual reality (VR) terminal device, augmented reality (AR) terminal device, wireless terminal device in industrial control, wireless terminal device in self-driving, wireless terminal device in remote medical surgery, wireless terminal device in smart grid, wireless terminal device in transportation safety, wireless terminal device in smart city, and wireless terminal device in smart home.

[0142] In some embodiments, the access network device is, for example, a node or device that connects a terminal to a wireless network. The access network device may include, but is not limited to, at least one of the following in a 5G communication system: evolved Node B (eNB), next-generation eNB (ng-eNB), next-generation Node B (gNB), node B (NB), home node B (HNB), home evolved node B (HeNB), radio backhaul device, radio network controller (RNC), base station controller (BSC), base transceiver station (BTS), base band unit (BBU), mobile switching center, base station in a 6G communication system, open RAN, cloud RAN, base station in other communication systems, and access node in a Wi-Fi system.

[0143] In some embodiments, the technical solutions of this disclosure can be applied to the Open RAN architecture. In this case, the interfaces between or within access network devices involved in the embodiments of this disclosure can be transformed into internal interfaces of Open RAN. The processes and information interactions between these internal interfaces can be implemented by software or programs.

[0144] In some embodiments, the access network device may be composed of a central unit (CU) and a distributed unit (DU). The CU may also be called a control unit. The CU-DU structure can separate the protocol layer of the access network device. Some of the protocol layer functions are centrally controlled by the CU, while the remaining part or all of the protocol layer functions are distributed in the DU and centrally controlled by the CU. However, this is not the only possibility.

[0145] In some embodiments, the core network equipment can be a single device, including a first network element, a second network element, etc., or it can be multiple devices or a group of devices, each including all or part of the first network element, the second network element, etc. Network elements can be virtual or physical. The core network includes, for example, at least one of the Evolved Packet Core (EPC), 5G Core Network (5GCN), and Next Generation Core (NGC).

[0146] It is understood that the communication system described in this disclosure is for the purpose of more clearly illustrating the technical solutions of this disclosure, and does not constitute a limitation on the technical solutions proposed in this disclosure. As those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions proposed in this disclosure are also applicable to similar technical problems.

[0147] The following embodiments of this disclosure can be applied to the communication system 100 shown in FIG1A, or to some of the main bodies, but are not limited thereto. The main bodies shown in FIG1A are illustrative. The communication system may include all or some of the main bodies in FIG1A, or it may include other main bodies outside of FIG1A. The number and form of each main body are arbitrary. Each main body may be physical or virtual. The connection relationship between the main bodies is illustrative. The main bodies may not be connected or may be connected. The connection can be in any way, it can be a direct connection or an indirect connection, it can be a wired connection or a wireless connection.

[0148] The embodiments disclosed herein can be applied to Long Term Evolution (LTE), LTE-Advanced (LTE-A), LTE-Beyond (LTE-B), SUPER 3G, IMT-Advanced, 4th generation mobile communication system (4G), 5th generation mobile communication system (5G), 5G new radio (NR), Future Radio Access (FRA), New-Radio Access Technology (RAT), New Radio (NR), New radio access (NX), Future generation radio access (FX), Global System for Mobile communications (GSM), CDMA2000, Ultra Mobile Broadband (UMB), IEEE 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), and IEEE 802.20, Ultra-Wideband (UWB), Bluetooth (a registered trademark), Public Land Mobile Network (PLMN) networks, Device-to-Device (D2D) systems, Machine-to-Machine (M2M) systems, Internet of Things (IoT) systems, Vehicle-to-Everything (V2X) systems, systems utilizing other communication methods, and next-generation systems built upon them, etc. Furthermore, multiple systems can be combined (e.g., a combination of LTE or LTE-A with 5G).

[0149] With the development of artificial intelligence (AI) and machine learning (ML) technologies, AI technology can reduce the feedback overhead of terminals or improve the accuracy of CSI feedback. For example, a bilateral AI / ML model can be used to generate a partial model based on the terminal-side channel state information (CSI) and a partial model based on the network-side CSI recovery model to achieve compressed feedback and recovery of CSI respectively.

[0150] Figure 1B is a schematic diagram illustrating CSI compression feedback and recovery based on a bilateral AI / ML model according to an embodiment of this disclosure. As shown in Figure 1B, the UE side compresses the downlink channel information H through a CSI generation partial model (e.g., called an encoder) and sends it to the network (NW) side device gNB after quantizing it into a binary bit stream. The gNB side recovers the downlink channel H', which is approximately the same as the original downlink information, through a CSI recovery partial model (defined as a decoder).

[0151] In some embodiments, the CSI generation and recovery partial models need to be trained using a collected dataset. The types of training methods for the CSI generation and recovery partial models include the following three:

[0152] Training Type 1: The model is trained on one side (such as the terminal side or the network side), and then the trained part of the model is sent to the other side;

[0153] Training Type 2: Train the CSI generation model and the CSI recovery model separately on the terminal side and the network side through joint training. Alternatively, after training a part of the model on one side of the terminal or network side, train another part of the bilateral model on the other side, where the parameters of the model trained first are not updated.

[0154] Training Type 3: First, the model is trained on one side, and then the training data or other auxiliary information is sent to the other side to train another part of the model.

[0155] In some embodiments, training Type 3 can be further divided into:

[0156] NW-first training: The network side first trains the CSI generation part model and the CSI recovery part model, and then sends the dataset used to train the CSI generation part model and / or other auxiliary information to the terminal side.

[0157] UE-first training: The terminal side first trains the CSI generation part model and the CSI recovery part model, and then sends the dataset used to train the CSI recovery part model and / or other auxiliary information to the network side.

[0158] In some embodiments, to reduce or mitigate the complexity of bilateral model training, the following inter-device collaborative training options can be employed:

[0159] Option 1: Standardize the model structure and parameters;

[0160] Option 2: Dataset standardization;

[0161] Option 3: Standardize the model structure, and the model parameters are passed between the NW side and the UE side;

[0162] Option 4: Standardize the data format, and the data is transmitted between the NW side and the UE side;

[0163] Option 5: Standardize the model format, while the reference model is passed between the NW side and the UE side.

[0164] In some embodiments, Option1 can be divided into:

[0165] Option 1-1: Standardized encoder;

[0166] Option 1-2: Standardized decoder;

[0167] Option 1-3: Standardize the encoder and decoder.

[0168] In some embodiments, based on the model parameters transmitted to the UE or the UE's execution behavior after modeling, Option 3 and Option 5 can be further divided into:

[0169] For Option 3, the following are included:

[0170] Option 3a: Receive model parameters and retrain the model to develop a different model or test it;

[0171] Option 3a-1: Pass the encoder parameter;

[0172] Option 3a-2: Pass parameters to the decoder;

[0173] Option 3a-3: Pass the encoder and decoder parameters;

[0174] Option 3b: The received model parameters are used directly for model inference.

[0175] For Option 5, the following are included:

[0176] Option 5a: Train the received model to develop a different model or test it;

[0177] Option 5a-1: Transmit the encoder model;

[0178] Option 5a-2: Transmit decoder model;

[0179] Option 5a-3: Pass the encoder and decoder models.

[0180] In some embodiments, depending on the content of the transmitted data, Option 4 can be further divided into:

[0181] For Option 4, the dataset is transferred from the NW side to the UE side, including:

[0182] Option 4-1: The dataset consists of target CSI and CSI feedback;

[0183] Option 4-2: The dataset consists of CSI feedback and the reconstructed target CSI;

[0184] Option 4-3: The dataset consists of target CSI, CSI feedback, and reconstructed target CSI.

[0185] In some embodiments, neural network pruning is a model optimization method aimed at reducing the complexity of the neural network, improving its efficiency, and preserving its performance as much as possible. Pruning simplifies the model by removing elements such as weights, neurons, or entire convolutional kernels. Here are some possible pruning methods:

[0186] Weight pruning: Removing weights from a neural network that have a relatively small impact on the output.

[0187] Neuron pruning: Removing an entire neuron, which typically involves removing neurons whose activation values ​​are consistently low or zero.

[0188] Kernel pruning: In convolutional neural networks (CNNs), the entire convolutional kernel (or filter) is removed.

[0189] Structural pruning: Removes an entire network layer or connections between network layers.

[0190] Channel pruning: In a multi-channel convolutional network, remove an entire channel.

[0191] The pruning methods described above can reduce model size, increase inference speed, prevent overfitting, or improve energy efficiency.

[0192] In some embodiments, when multiple AI models are deployed on the terminal side, transmitting the parameters of multiple models will consume a large amount of wireless resources. How to reduce the resource overhead caused by transmitting model parameters is a problem to be solved.

[0193] Figure 2 is an interactive schematic diagram of a communication method according to an embodiment of the present disclosure. As shown in Figure 2, the embodiments of the present disclosure relate to a communication method, which includes:

[0194] Step S2101: The terminal sends the first information to the network device.

[0195] In some embodiments, the first information is used to indicate a first model structure. Optionally, the first model structure is a model structure for which the terminal needs the network device to transmit model parameters.

[0196] In some embodiments, the first model structure may include a model structure for which the terminal has not configured model parameters, or a model structure for which the model parameters need to be updated. Optionally, the first model structure may include one or more of the following: model structures supported by the terminal, model structures available to the terminal, and model structures that are already activated on the terminal.

[0197] In some embodiments, the number of first model structures can be one or more.

[0198] In some embodiments, the terminal may have one or more model structures pre-deployed. For example, the number and connection relationships of each sub-model structure in the model may be pre-determined, such as the number of neural network layers, the number of neurons, the number of channels, the connection relationships of neurons, the connection relationships of neural network layers, etc. The specific model parameters in these model structures may be unknown to the terminal. Therefore, the terminal can inform the network device of the corresponding model structure through the first information so that the network device can transmit the corresponding model parameters to the terminal.

[0199] In some embodiments, the above model structure can be applied to, but is not limited to, AI-based CSI enhancement, AI-based beam management, and AI-based positioning.

[0200] In some embodiments, the first model structure may be a model structure of a CSI-generated partial model (such as an encoder), a model structure of a CSI-recovered partial model (such as a decoder), or a model structure of any other model. This disclosure does not limit this type of model.

[0201] In some embodiments, each model structure may have a unique corresponding identifier. Optionally, for a bilateral model, a set of encoders and decoders may have the same model identifier or a pair identifier. Optionally, each encoder has its corresponding encoder identifier, and each decoder has its corresponding decoder identifier, with each encoder identifier uniquely corresponding to a decoder identifier.

[0202] In some embodiments, the first information includes an identifier corresponding to each first model structure. Optionally, the representation corresponding to the first model structure can be any of the following: model identifier; encoder identifier; decoder identifier; pairing identifier.

[0203] In some embodiments, the first information may also be referred to as "parameter request information", "structural indication information", "capability information", etc., and the names are not limited in the embodiments disclosed herein.

[0204] In some embodiments, the network device receives first information. Optionally, the network device determines a first model structure based on the first information. Optionally, the network device executes step S2102 in response to the first information.

[0205] In step S2102, the network device sends the second information to the terminal.

[0206] In some embodiments, the second information is used to indicate the model parameters of the first model, and the model structure of the first model is the first model structure.

[0207] In some embodiments, the number of first models can be one or more. Optionally, the second information is used to indicate the model parameters corresponding to all first model structures, or the model parameters corresponding to some first model structures.

[0208] For example, the first information indicates four model structures. The network device can indicate the model parameters corresponding to these four model structures through the second information, or it can indicate only the model parameters of some model structures.

[0209] In some embodiments, the model parameters of the first model include at least one model parameter corresponding to the first model structure, and the second information is also used to indicate the model structure corresponding to each of the at least one model parameter.

[0210] For example, if the first information sent by the terminal indicates four model structures, the network device can indicate only the model parameters corresponding to two of the model structures, i.e., two sets of model parameters. At the same time, the network device can also indicate which model structure each of these two sets of model parameters corresponds to, so as to ensure that the terminal can accurately configure the model parameters of the model.

[0211] In some embodiments, the model parameters of the first model include only the model parameters corresponding to the first sub-model structure. The first sub-model structure includes at least one of the following: a portion of the neural network layer in the first model, a portion of the neurons in the first model, a portion of the channels in the first model, and a portion of the convolutional kernels in the first model.

[0212] It is understandable that any model structure can include multiple sub-model structures. For example, a model structure can include multiple neural network layers, multiple neurons, and multiple channels, or, for a convolutional neural network model, it can include multiple convolutional kernels.

[0213] In some embodiments, the first sub-model structure may be an unstandardized or undefined sub-model structure in each first model.

[0214] Optionally, for some model structures, the model parameters of some sub-model structures can be predefined. In this case, the network device may not need to transmit the model parameters of these predefined sub-model structures to the terminal. For example, if a model structure includes four neural network layers, and the model parameters of two of the neural network layers are predefined, the second information can only indicate the model parameters corresponding to the other two neural network layers in the model structure.

[0215] In some embodiments, the model parameters of the first model can be pruned model parameters. That is, the network device can prune the trained model and transmit the pruned model parameters to the terminal.

[0216] In some embodiments, the network device may prune the trained model using one or more of the following pruning methods: weight pruning, neuron pruning, convolution kernel pruning, structural pruning, and channel pruning.

[0217] In some embodiments, the network device may indicate the pruning method. Optionally, the network device may also indicate the content that was pruned. Optionally, the second information is also used to indicate the pruning method applied to the pruned content and / or the model parameters of the first model.

[0218] In some embodiments, the network device can indicate the pruning method of the pruned content and / or the model parameters of the first model through the second information, and can also indicate the structure of the second sub-model through other information. For example, the network device can send third information to the terminal separately to indicate the pruning method of the pruned content and / or the model parameters of the first model.

[0219] In some embodiments, the model parameters of the first model are subject to weight pruning. The pruning includes the first neuron and / or the first neural network layer. The first neuron is the neuron whose model parameters are pruned in the first model, and the first neural network layer is the neural network layer whose model parameters are pruned in the first model.

[0220] The model parameters of the first model are pruned by neurons. The pruning includes the first neuron, which is the neuron in the first model whose corresponding model parameters are pruned.

[0221] The model parameters of the first model are pruned by convolution kernels. The pruning includes the first convolution kernel, which is the convolution kernel in the first model whose corresponding model parameters are pruned.

[0222] The model parameters of the first model are structurally pruned. The pruning includes the first neural network layer and / or the second model parameters. The first neural network layer is the neural network layer in the first model whose corresponding model parameters are pruned. The second model parameters are the model parameters between the neural network layers in the first model that are pruned.

[0223] The model parameters of the first model are pruned through channels. The pruning includes the first channel, which is the channel in the first model where the corresponding model parameters are pruned.

[0224] In some embodiments, the second information is used to indicate the model parameters corresponding to the unpruned sub-model structures. For example, if a model has undergone neuron pruning and channel pruning, the second information can only indicate the model parameters corresponding to the unpruned neurons and the unpruned channels, without transmitting the pruned content. Accordingly, the second information can indicate which neurons and which channels have been pruned, so that the terminal can more accurately determine the model parameters.

[0225] In some embodiments, the model parameters of the first model are quantized model parameters. Optionally, the neural network layers or neurons of the first model are divided into multiple parts, and the model parameters corresponding to each part correspond to different quantization precisions.

[0226] Optionally, quantization precision can be represented by the number of bits corresponding to quantization. For example, the higher the number of bits corresponding to quantization, the higher the quantization precision. For instance, if a part of the model parameters is quantized using 2 bits, then a data point can only be quantized into 4 different values. If 3 bits are used for quantization, then a data point can be quantized into 8 different values.

[0227] For example, a neural network layer in a model can be divided into N parts. More important parts can use higher bit counts for quantization, while less important parts can use lower bit counts. For instance, if the parameters of a certain neural network layer are more important, 3-bit quantization can be used, while the parameters of other neural network layers or neurons can be quantized using 2-bit quantization.

[0228] In some embodiments, some sub-models with different structures may have the same model parameters. For example, some neural network layers in multiple models may have the same structure and parameters, or some neurons in multiple models may have the same structure and parameters, or some channels in multiple models may have the same structure and parameters, or some convolutional kernels in multiple convolutional network models may have the same structure and parameters.

[0229] In some embodiments, the first model includes multiple models with different model structures, and the multiple models include a second sub-model structure with the same model structure. The model parameters corresponding to the second sub-model structure are the same, and the second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

[0230] For example, a network device can transmit model parameters corresponding to model A and model B to a terminal. These two models have different model structures, but they can have some identical sub-model structures, and these sub-model structures can use the same model parameters. For example, both models include two neural network layers with the same structure, and the parameters corresponding to these two neural network layers are also the same. In this case, since the structure and parameters of these two neural network layers in model A and model B are the same, the second information can only include the model parameters corresponding to the two neural network layers in model A, without transmitting the model parameters corresponding to the two neural network layers in model B.

[0231] In some embodiments, the second information is further used to indicate the second sub-model structure; or, the terminal negotiates with the network device to determine the second sub-model structure; or, the terminal reports the second sub-model structure.

[0232] In other words, determining which sub-model structures among multiple models share the same model parameters can be accomplished by the network device, through negotiation, or by the terminal and reported. This allows the network device and the terminal to reach a consensus, ensuring the accuracy of the model parameter configuration.

[0233] Alternatively, the model parameters corresponding to which sub-model solution structures in multiple models are different can be determined and indicated by the network device, determined through negotiation, or determined and reported by the terminal.

[0234] In some embodiments, the network device can indicate the second sub-model structure not only through the second information, but also through other information. For example, the network device can send a fourth piece of information to the terminal to indicate the second sub-model structure.

[0235] In some embodiments, when the transport rank is greater than 1, the structural parameters of some or all data transmission layers can be designed to be the same. For example, two of the four transport layers use the same CSI generation partial model for CSI enhancement.

[0236] In some embodiments, when the transmission rank is greater than 1, it is possible to design that some or all of the transmission layers have sub-model structures with the same structure and parameters. For example, in the CSI generation sub-models of two of the four transmission layers, one or more neural network layers have the same structure and parameters.

[0237] In some embodiments, the transmission rank is greater than 1, and the model parameters of the first model satisfy at least one of the following: the model parameters of the models corresponding to multiple or all data transmission layers are the same; the models corresponding to multiple or all data transmission layers include a second sub-model structure with the same model structure, and the model parameters of the second sub-model structure are the same; wherein, the second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; a convolutional kernel.

[0238] For example, if the model parameters of the models corresponding to transport layer A and transport layer B are the same, the network device can transmit only the model parameters of the model corresponding to transport layer A to the terminal, without transmitting the model parameters of the model corresponding to transport layer B. Alternatively, if the models corresponding to transport layer A and transport layer B each include the same neural network layer, the network device can transmit only the parameters of that neural network layer in the model corresponding to transport layer A to the terminal, without transmitting the parameters of that neural network layer in the model corresponding to transport layer B.

[0239] In some embodiments, the network device may indicate which data transmission layers have the same model parameters. Optionally, the network device may indicate which sub-model structures have the same model parameters among the models corresponding to multiple or all data transmission layers.

[0240] Alternatively, the network device can indicate which data transmission layers have different model parameters. Optionally, the network device can indicate which sub-model structures have different model parameters among the models corresponding to multiple or all data transmission layers.

[0241] In some embodiments, which data transmission layers have the same (or different) model parameters, and / or which sub-model structures have the same (or different) model parameters, can be determined and indicated by the network device. Optionally, the network device indicates which data transmission layers have the same (or different) model parameters, and / or which sub-model structures have the same (or different) model parameters, through second information. Optionally, the second information is used to indicate data transmission layers with the same (or different) model parameters. Optionally, the network device may also indicate data transmission layers with the same model parameters through other information, such as sending fifth information separately to indicate which data transmission layers have the same (or different) model parameters.

[0242] In some embodiments, the terminal and the network device negotiate to determine a data transmission layer with the same (or different) model parameters, and / or a second sub-model structure (or a sub-model structure other than the second sub-model structure). Optionally, the terminal reports a data transmission layer with the same (or different) model parameters, and / or a second sub-model structure (or a sub-model structure other than the second sub-model structure).

[0243] In some embodiments, the second information is further used to indicate a second model structure, which is a model structure for which the second information in the first model structure does not indicate the corresponding model parameters.

[0244] For example, if the first information sent by the terminal indicates four model structures, the network device may indicate only the model parameters corresponding to two of the model structures. At the same time, the network device may also indicate which two model structures among the four model structures did not have their model parameters indicated by the second information.

[0245] In some embodiments, the second information may be, for example, Radio Resource Control (RRC), Media Access Control Control Element (MAC-CE), or Downlink Control Information (DCI) signaling, but is not limited thereto.

[0246] In some embodiments, the second information may also be referred to as "parameter indication information", "model parameter configuration", etc., and this disclosure does not limit it.

[0247] In some embodiments, the names of information, etc., are not limited to the names described in the embodiments. Terms such as "information", "message", "signal", "signaling", "report", "configuration", "indication", "instruction", "command", "channel", "parameter", "domain", "field", "symbol", "symbol", "codebook", "codeword", "codepoint", "bit", "data", "program", and "chip" can be used interchangeably.

[0248] In some embodiments, "acquire," "get," "obtain," "receive," "transmit," "bidirectional transmission," and "send and / or receive" can be used interchangeably and can be interpreted as receiving from other entities, acquiring from protocols, acquiring from higher layers, obtaining through self-processing, or autonomous implementation. Protocols include, for example, at least one of the 3GPP protocol, Wi-Fi protocol, and audio and / or video protocols.

[0249] In some embodiments, terms such as “send,” “transmit,” “report,” “distribute,” “transfer,” “bidirectional transmission,” “send and / or receive” can be used interchangeably.

[0250] In some embodiments, terms such as "certain," "preset," "default," "set," "indicated," "a certain," "any," and "first" can be used interchangeably. "Certain A," "preset A," "default A," "set A," "indicated A," "a certain A," "any A," and "first A" can be interpreted as A pre-defined in a protocol or the like, or as A obtained through setting, configuration, or instruction, or as specific A, a certain A, any A, or first A, but are not limited thereto.

[0251] In some embodiments, the determination or judgment can be made by a value represented by 1 bit (0 or 1), or by a true or false value (boolean), or by a comparison of numerical values ​​(e.g., a comparison with a predetermined value), but is not limited thereto.

[0252] In some embodiments, "not expecting to receive" can be interpreted as not receiving on time domain resources and / or frequency domain resources, or as not performing subsequent processing on the data and / or instructions received; "not expecting to send" can be interpreted as not sending, or as sending but not expecting the receiver to respond to the sent content.

[0253] The communication method involved in the embodiments of this disclosure may include at least one of steps S2101 to S2102. For example, step S2101 may be implemented as a separate embodiment, and step S2102 may be implemented as a separate embodiment, but are not limited thereto.

[0254] In some embodiments, step S2101 is optional, and one or more of these steps may be omitted or substituted in different embodiments.

[0255] In some embodiments, step S2102 is optional, and one or more of these steps may be omitted or substituted in different embodiments.

[0256] In some embodiments, the steps and their optional implementations in other embodiments described before or after this embodiment, as well as other related parts in the specification, can be referred to, and will not be repeated here.

[0257] Figure 3A is a flowchart illustrating a communication method according to an embodiment of the present disclosure. As shown in Figure 3A, the present disclosure relates to a communication method that can be executed by a terminal, and the method includes:

[0258] Step S3101: Send the first message.

[0259] The optional implementation of step S3101 can be found in the optional implementation of step S2101 in Figure 2 and other related parts in the embodiments involved in Figure 2, which will not be repeated here.

[0260] Step S3102: Obtain the second information.

[0261] The optional implementation of step S3102 can be found in the optional implementation of step S2102 in Figure 2 and other related parts in the embodiments involved in Figure 2, which will not be repeated here.

[0262] In some embodiments, a first message is sent to the network device, the first message being used to indicate a first model structure, the first model structure being a model structure for which the terminal needs the network device to transmit model parameters;

[0263] The system receives second information sent by a network device. The second information is used to indicate the model parameters of the first model, and the model structure of the first model is the first model structure.

[0264] In some embodiments, the model parameters of the first model include only the model parameters corresponding to the first sub-model structure. The first sub-model structure includes at least one of the following: a portion of the neural network layer in the first model, a portion of the neurons in the first model, a portion of the channels in the first model, and a portion of the convolutional kernels in the first model.

[0265] In some embodiments, the model parameters of the first model are pruned model parameters.

[0266] In some embodiments, the second information is also used to indicate the pruning content;

[0267] Pruning includes at least one of the following:

[0268] The model parameters of the first model are pruned by weights. The pruning includes the first neuron and / or the first neural network layer. The first neuron is the neuron whose model parameters are pruned in the first model, and the first neural network layer is the neural network layer whose model parameters are pruned in the first model.

[0269] The model parameters of the first model are pruned by neurons. The pruning includes the first neuron, which is the neuron in the first model whose corresponding model parameters are pruned.

[0270] The model parameters of the first model are pruned by convolution kernels. The pruning includes the first convolution kernel, which is the convolution kernel in the first model whose corresponding model parameters are pruned.

[0271] The model parameters of the first model are structurally pruned. The pruning includes the first neural network layer and / or the second model parameters. The first neural network layer is the neural network layer in the first model whose corresponding model parameters are pruned. The second model parameters are the model parameters between the neural network layers in the first model that are pruned.

[0272] The model parameters of the first model are pruned through channels. The pruning includes the first channel, which is the channel in the first model where the corresponding model parameters are pruned.

[0273] In some embodiments, the model parameters of the first model are quantized model parameters;

[0274] In the first model, the neural network layer or neurons are divided into multiple parts, and the model parameters corresponding to each part correspond to different quantization precisions.

[0275] In some embodiments, the first model includes multiple models with different model structures, and the multiple models include a second sub-model structure with the same model structure. The model parameters corresponding to the second sub-model structure are the same, and the second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

[0276] In some embodiments, the second information is further used to indicate the structure of the second sub-model; or,

[0277] The terminal and network equipment negotiate to determine the structure of the second sub-model; or,

[0278] The terminal reports the structure of the second sub-model.

[0279] In some embodiments, the transmission rank is greater than 1, and the model parameters of the first model satisfy at least one of the following:

[0280] The model parameters of the models corresponding to multiple or all data transmission layers are the same;

[0281] The models corresponding to multiple or all data transmission layers include second sub-model structures with the same model structure and the same model parameters in the second sub-model structures.

[0282] The second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

[0283] In some embodiments, the second information is further used to indicate a data transmission layer with the same model parameters, and / or, a second sub-model structure; or,

[0284] The terminal and network device negotiate and determine a data transmission layer with the same model parameters, and / or, a second sub-model structure; or,

[0285] The terminal reports the same data transmission layer as the model parameters, and / or the second sub-model structure.

[0286] In some embodiments, the first information includes an identifier corresponding to each first model structure, which is any one of the following: model identifier, encoder identifier, decoder identifier, or pairing identifier.

[0287] In some embodiments, the model parameters of the first model include at least one model parameter corresponding to the first model structure, and the second information is also used to indicate the model structure corresponding to each of the at least one model parameter.

[0288] In some embodiments, the second information is further used to indicate a second model structure, which is a model structure for which the second information in the first model structure does not indicate the corresponding model parameters.

[0289] In some embodiments, the steps and their optional implementations in other embodiments described before or after this embodiment, as well as other related parts in the specification, can be referred to, and will not be repeated here.

[0290] Figure 3B is a flowchart illustrating a communication method according to an embodiment of the present disclosure. As shown in Figure 3B, the present disclosure relates to a communication method that can be executed by a network device, and the method includes:

[0291] Step S3201: Obtain the first information.

[0292] The optional implementation of step S3201 can be found in the optional implementation of step S2101 in Figure 2 and other related parts in the embodiments involved in Figure 2, which will not be repeated here.

[0293] Step S3202: Send the second message.

[0294] The optional implementation of step S3201 can be found in the optional implementation of step S2101 in Figure 2 and other related parts in the embodiments involved in Figure 2, which will not be repeated here.

[0295] In some embodiments, the receiving terminal sends first information, which is used to indicate a first model structure, and the first model structure is a model structure for which the terminal needs the network device to transmit model parameters.

[0296] Send a second message to the terminal. The second message is used to indicate the model parameters of the first model. The model structure of the first model is the first model structure.

[0297] In some embodiments, the model parameters of the first model include only the model parameters corresponding to the first sub-model structure. The first sub-model structure includes at least one of the following: a portion of the neural network layer in the first model, a portion of the neurons in the first model, a portion of the channels in the first model, and a portion of the convolutional kernels in the first model.

[0298] In some embodiments, the model parameters of the first model are pruned model parameters.

[0299] In some embodiments, the second information is also used to indicate the pruning content;

[0300] Pruning includes at least one of the following:

[0301] The model parameters of the first model are pruned by weights. The pruning includes the first neuron and / or the first neural network layer. The first neuron is the neuron whose model parameters are pruned in the first model, and the first neural network layer is the neural network layer whose model parameters are pruned in the first model.

[0302] The model parameters of the first model are pruned by neurons. The pruning includes the first neuron, which is the neuron in the first model whose corresponding model parameters are pruned.

[0303] The model parameters of the first model are pruned by convolution kernels. The pruning includes the first convolution kernel, which is the convolution kernel in the first model whose corresponding model parameters are pruned.

[0304] The model parameters of the first model are structurally pruned. The pruning includes the first neural network layer and / or the second model parameters. The first neural network layer is the neural network layer in the first model whose corresponding model parameters are pruned. The second model parameters are the model parameters between the neural network layers in the first model that are pruned.

[0305] The model parameters of the first model are pruned through channels. The pruning includes the first channel, which is the channel in the first model where the corresponding model parameters are pruned.

[0306] In some embodiments, the model parameters of the first model are quantized model parameters;

[0307] In the first model, the neural network layer or neurons are divided into multiple parts, and the model parameters corresponding to each part correspond to different quantization precisions.

[0308] In some embodiments, the first model includes multiple models with different model structures, and the multiple models include a second sub-model structure with the same model structure. The model parameters corresponding to the second sub-model structure are the same, and the second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

[0309] In some embodiments, the second information is further used to indicate the structure of the second sub-model; or,

[0310] The terminal and network equipment negotiate to determine the structure of the second sub-model; or,

[0311] The terminal reports the structure of the second sub-model.

[0312] In some embodiments, the transmission rank is greater than 1, and the model parameters of the first model satisfy at least one of the following:

[0313] The model parameters of the models corresponding to multiple or all data transmission layers are the same;

[0314] The models corresponding to multiple or all data transmission layers include second sub-model structures with the same model structure and the same model parameters in the second sub-model structures.

[0315] The second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

[0316] In some embodiments, the second information is further used to indicate a data transmission layer with the same model parameters, and / or, a second sub-model structure; or,

[0317] The terminal and network device negotiate and determine a data transmission layer with the same model parameters, and / or, a second sub-model structure; or,

[0318] The terminal reports the same data transmission layer as the model parameters, and / or the second sub-model structure.

[0319] In some embodiments, the first information includes an identifier corresponding to each first model structure, which is any one of the following: model identifier, encoder identifier, decoder identifier, or pairing identifier.

[0320] In some embodiments, the model parameters of the first model include at least one model parameter corresponding to the first model structure, and the second information is also used to indicate the model structure corresponding to each of the at least one model parameter.

[0321] In some embodiments, the second information is further used to indicate a second model structure, which is a model structure for which the second information in the first model structure does not indicate the corresponding model parameters.

[0322] In some embodiments, the steps and their optional implementations in other embodiments described before or after this embodiment, as well as other related parts in the specification, can be referred to, and will not be repeated here.

[0323] Figure 4A is an interactive schematic diagram illustrating a communication method according to an embodiment of the present disclosure. As shown in Figure 4A, the present disclosure relates to a communication method, which includes:

[0324] Step S4101: The UE sends the ID (IDentity) associated with the model structure to the NW.

[0325] In some embodiments, the UE can report the supported, available, activated, or NW-required model structures to the NW via capability reporting. Each model structure can be associated with an ID, which can be a model ID, encoder ID, decoder ID, or pairing ID, etc. The UE sends the indication information of the model structure associated ID to the NW.

[0326] In step S4102, the NW sends the model parameters and / or the ID associated with the model structure to the UE.

[0327] In some embodiments, if the NW side contains model parameters corresponding to the reported ID, the NW sends one or more model parameters corresponding to the included ID to the UE side after quantization. When the NW side contains parameters corresponding to multiple model structures, and the NW only transmits the parameters of one model, it also needs to send the model ID corresponding to the transmitted model parameter and the model parameter together to the UE. If the NW side does not contain model parameters corresponding to the reported ID, the NW sends a message to the UE indicating that there are no model parameters with the corresponding ID to be transmitted.

[0328] In some embodiments, it is assumed that the UE supports four model structures, whose associated mode IDs are defined as ID#1, ID#2, ID#3, and ID#4, respectively. However, due to application limitations on the UE side, it may ultimately only support the two model structures ID#1 and ID#3. In step S4101, the UE can send the ID#1 and ID#3 corresponding to the two supported model structures to the NW through capability reporting.

[0329] In some embodiments, if the NW side can determine, based on the current working environment or the application conditions of the NW side, that it only needs to transmit the model parameters corresponding to ID#3 to the UE in step S4102, the NW may also send the indication information of ID#3 to the UE so that the UE can know that the model parameters transmitted in step S4102 are the model parameters corresponding to ID#3.

[0330] Optionally, the NW only sends the configuration information corresponding to ID#3 to the UE, without including the indication information of ID#3. The UE determines which model structure the NW is transmitting the model parameters based on the NW's configuration information.

[0331] In some embodiments, the parameters of some neural network layers of the encoder or decoder model can be standardized, and NW only transmits the parameters corresponding to the unstandardized neural network layers.

[0332] In some embodiments, model parameters can be passed on after model pruning.

[0333] Optionally, the NW prunes the trained AI model and passes the pruned model parameters to the UE. Furthermore, the NW sends pruning instructions to the UE. These instructions include at least the pruning method and the specific pruning content. Different pruning methods correspond to different pruning content.

[0334] For weight pruning, the pruning content can be used to indicate which neural network layers, neurons, and other corresponding model parameters are removed.

[0335] For neuron pruning, the pruning parameters can be used to indicate which model parameters corresponding to which neurons are removed.

[0336] For convolutional kernel pruning, the pruning content can be used to: in a convolutional neural network, indicate that the model parameters corresponding to the convolutional kernel are removed.

[0337] For structural pruning, the pruning content can be used to indicate which neural network layers or the corresponding model parameters between neural network layers are removed.

[0338] For channel pruning, the pruning content can be used to indicate which channel's corresponding model parameters are removed.

[0339] In some embodiments, uniform or non-uniform model parameter quantization can be performed.

[0340] Optionally, the neural network layer or neural unit in the model can be divided into N parts, and the model parameters corresponding to the nth part can be quantized using Xn bits, where Xn can represent the number of bits corresponding to the nth part, and Xi≠Xj.

[0341] In some embodiments, if the UE deploys multiple AI models with different structures, the following design method can be used to reduce the transmission of model parameters between different AI models.

[0342] Optionally, the model parameters corresponding to some neural network information in multiple AI models with different structures are the same. The partial neural network information includes one or more of the following: partial neural network layers, partial neurons, partial channels, and for convolutional networks, partial convolutional kernels.

[0343] In some embodiments, for Rank>1, the following design method is used to reduce the transmission of model parameters corresponding to different data transmission layers.

[0344] Optionally, the model parameters corresponding to some or all of the data transmission layers can be the same.

[0345] Optionally, the model parameters of some or all of the data transmission layers are the same for the corresponding neural network layers, neurons, channels, or convolutional kernels.

[0346] In some embodiments, the model parameters of which data transmission layers are the same, or the model parameters of which parts of the neural network layers are the same, are determined by the following methods: the NW sends a message indication to the UE; the NW and the UE negotiate predefined parameters; and the UE sends a message to the NW.

[0347] In some embodiments, assuming the encoder model uses the model structure shown in Figure 4B, it is standardized, and only the model parameters corresponding to each neural network layer in the ResNet block in Figure 4B are standardized. Other model parameters are updated on the NW side, and the NW can send them to the UE through step S4102. Alternatively, the model parameters corresponding to each neural network layer in the ResNet block or some blocks in the ResNet block are not standardized, while the remaining model parameters are standardized. In step S4102, the NW can only transmit the model parameters corresponding to the ResNet block to the UE.

[0348] In some embodiments, NW removes certain convolutional kernels from the trained encoder, or sets the parameters corresponding to these kernels to 0. In this case, NW does not need to transmit the pruned kernel parameters, thus reducing the overhead of transmitting model parameters. Similarly, NW may also prune parameters corresponding to certain structures or channels.

[0349] In some embodiments, the model parameters corresponding to neural network layers may have varying degrees of importance, with some layer weights being less critical and others more critical. These weights are quantized using different bit values ​​based on their importance. For example, referring to Figure 4B, the parameters corresponding to ResNet blocks are more critical, and each parameter can be quantized using 3 bits. The model parameters corresponding to the remaining network layers or neurons are quantized using 2 bits. This quantization method reduces the overhead of transmitting model parameters while maintaining nearly identical performance.

[0350] In some embodiments, it is assumed that three encoder models are trained on the NW side, corresponding to the UMA, UMi, and indoor scenes respectively, and the model structures of these three encoders are identical. Assuming that the NW has already completed the training of the encoder model for the UMA scene, when training the encoders for the UMi and indoor scenes, the parameters of some neural network layers can remain unchanged, and only the parameters of other neural network layers need to be updated.

[0351] Among them, UMA scenario can be a macro cell scenario in urban areas, where the user density is high both indoors and outdoors, and the base station is higher than the surrounding buildings. UMi scenario can be a micro cell scenario in urban areas, where the user density is high both indoors and outdoors, and the base station is lower than the surrounding buildings. Indoor scenario can be an indoor scenario.

[0352] Figure 4B is a schematic diagram of an encoder model structure according to an embodiment of the present disclosure. As shown in Figure 4B, the model structure includes an input layer, a convolutional layer, a 2D batch normalization and corrected linear unit (ReLU) layer, a residual (ResNet) layer, a convolutional layer, a reshape layer, a fully connected (FC) layer, a scalar quantization layer, and an output layer connected in sequence.

[0353] The residual layer can include six ResNet blocks. Each residual block can include a convolutional layer, a 2D batch normalization and rectified linear unit layer, another convolutional layer, a 2D batch normalization layer, and a rectified linear unit layer, all connected in sequence. The scalar quantization layer can include 32 neurons A, which can convert a floating-point number into 2-bit data. Optionally, the input to this model structure can be two samples, each sample including 13 channels of features, and each channel of features including 32 real-valued data, which can be floating-point data. The output of this model structure can be 64-bit data.

[0354] Referring to Figure 4B, if the parameters of the ResNet block need to be changed while the parameters of other parts remain unchanged, and if the NW side has already passed the corresponding model parameters for the UMA scenario to the UE, then in the UMi or indoor scenario, the NW side only needs to pass the parameters of the RestNet block corresponding to the encoder model of the UMi or indoor scenario to the UE. The parameters of other parts do not need to be passed again.

[0355] As for which parameters of a neural network layer, neuron, channel, or convolution kernel are variable, and which parameters are constant, the NW side can send the indication information of the variable or constant neural network layer, neuron, channel, or convolution kernel to the UE through one or more of RRC, MAC-CE, or DCI signaling.

[0356] In some embodiments, when the transmission rank is greater than 1, the model parameters corresponding to the encoders of different data transmission layers or some data transmission layers are the same, or the parameters corresponding to the neurons, channels, or convolutional kernels of some neural network layers corresponding to the encoders of different layers are the same. The NW side still indicates to the UE which data layers have the same model parameters, or which data transmission layers have the same model parameters corresponding to some neural network layers, neurons, channels, or convolutional kernels, through RRC, MAC-CE, or DCI signaling.

[0357] In the above embodiments, the resource overhead caused by model parameter passing can be reduced, and system performance can be improved.

[0358] In the embodiments disclosed herein, some or all of the steps and their optional implementations may be arbitrarily combined with some or all of the steps in other embodiments, or may be arbitrarily combined with the optional implementations in other embodiments.

[0359] This disclosure also provides an apparatus for implementing any of the above methods. For example, an apparatus is provided that includes units or modules for implementing the steps performed by the terminal in any of the above methods. Alternatively, another apparatus is provided that includes units or modules for implementing the steps performed by a network device (e.g., an access network device, a core network functional node, a core network device, etc.) in any of the above methods.

[0360] It should be understood that the division of units or modules in the above device is only a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, the units or modules in the device can be implemented by a processor calling software: for example, the device includes a processor connected to a memory containing instructions. The processor calls the instructions stored in the memory to implement any of the above methods or to implement the functions of the units or modules in the above device. The processor can be, for example, a general-purpose processor, such as a Central Processing Unit (CPU) or a microprocessor, and the memory can be internal or external to the device. Alternatively, the units or modules in the device can be implemented in the form of hardware circuits. The functionality of some or all of the units or modules can be achieved through the design of these hardware circuits, which can be understood as one or more processors. For example, in one implementation, the hardware circuit is an application-specific integrated circuit (ASIC). The functionality of some or all of the units or modules is achieved through the design of the logical relationships between the components within the circuit. In another implementation, the hardware circuit can be implemented using a programmable logic device (PLD). Taking a field-programmable gate array (FPGA) as an example, it can include a large number of logic gates. The connection relationships between the logic gates are configured through configuration files, thereby achieving the functionality of some or all of the units or modules. All units or modules of the above device can be implemented entirely through processor-called software, entirely through hardware circuits, or partially through processor-called software with the remaining parts implemented through hardware circuits.

[0361] In this embodiment, the processor is a circuit with signal processing capabilities. In one implementation, the processor can be a circuit with instruction read and execute capabilities, such as a Central Processing Unit (CPU), a microprocessor, a graphics processing unit (GPU) (which can be understood as a microprocessor), or a digital signal processor (DSP). In another implementation, the processor can implement certain functions through the logical relationships of hardware circuits. The logical relationships of the aforementioned hardware circuits are fixed or reconfigurable. For example, the processor is a hardware circuit implemented using an application-specific integrated circuit (ASIC) or a programmable logic device (PLD), such as an FPGA. In a reconfigurable hardware circuit, the process of the processor loading a configuration document and configuring the hardware circuit can be understood as the process of the processor loading instructions to implement the functions of some or all of the above units or modules. Furthermore, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as a Neural Network Processing Unit (NPU), a Tensor Processing Unit (TPU), or a Deep Learning Processing Unit (DPU).

[0362] Figure 5A is a schematic diagram of the structure of a terminal proposed in an embodiment of this disclosure. As shown in Figure 5A, the terminal 5100 may include at least one of a transceiver module 5101, a processing module 5102, etc.

[0363] In some embodiments, the transceiver module 5101 is configured to send first information to a network device, the first information being used to indicate a first model structure, the first model structure being a model structure for which the terminal needs the network device to transmit model parameters;

[0364] The transceiver module 5101 is further configured to receive second information sent by the network device, the second information being used to indicate the model parameters of the first model, and the model structure of the first model being the first model structure.

[0365] Optionally, the transceiver module 5101 is used to perform at least one of the communication steps such as sending and / or receiving performed by the terminal in any of the above methods, which will not be described in detail here. Optionally, the processing module 5102 is used to perform at least one of the other steps performed by the terminal in any of the above methods, which will not be described in detail here.

[0366] Figure 5B is a schematic diagram of the structure of a network device proposed in an embodiment of this disclosure. As shown in Figure 5B, the network device 5200 may include at least one of a transceiver module 5201, a processing module 5202, etc.

[0367] In some embodiments, the transceiver module 5201 is configured to receive first information sent by the terminal, the first information being used to indicate a first model structure, the first model structure being a model structure for which the terminal needs the network device to transmit model parameters;

[0368] The transceiver module 5201 is also used to send second information to the terminal, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

[0369] Optionally, the transceiver module 5201 is used to perform at least one of the communication steps such as sending and / or receiving performed by the network device in any of the above methods, which will not be described in detail here. Optionally, the processing module 5202 is used to perform at least one of the other steps performed by the network device in any of the above methods, which will not be described in detail here.

[0370] In some embodiments, the transceiver module may include a transmitting module and / or a receiving module, which may be separate or integrated. Optionally, the transceiver module may be interchangeable with a transceiver.

[0371] In some embodiments, the processing module may be a single module or may include multiple sub-modules. Optionally, the multiple sub-modules may each perform all or part of the steps required by the processing module. Optionally, the processing module may be interchangeable with a processor.

[0372] Figure 6A is a schematic diagram of the structure of the communication device 6100 proposed in an embodiment of this disclosure. The communication device 6100 can be a network device (e.g., access network device, core network device, etc.), a terminal (e.g., user equipment, etc.), a chip, chip system, or processor that supports the network device in implementing any of the above methods, or a chip, chip system, or processor that supports the terminal in implementing any of the above methods. The communication device 6100 can be used to implement the methods described in the above method embodiments; for details, please refer to the descriptions in the above method embodiments.

[0373] As shown in Figure 6A, the communication device 6100 includes one or more processors 6101. The processor 6101 can be a general-purpose processor or a dedicated processor, such as a baseband processor or a central processing unit (CPU). The baseband processor can be used to process communication protocols and communication data, while the CPU can be used to control communication devices (e.g., base stations, baseband chips, terminal devices, terminal device chips, DUs or CUs, etc.), execute programs, and process program data. Optionally, the communication device 6100 can be used to execute any of the above methods. Optionally, one or more processors 6101 can be used to invoke instructions to cause the communication device 6100 to execute any of the above methods.

[0374] In some embodiments, the communication device 6100 further includes one or more transceivers 6102. When the communication device 6100 includes one or more transceivers 6102, the transceiver 6102 performs at least one of the communication steps such as sending and / or receiving in the above method, and the processor 6101 performs at least one of the other steps. In optional embodiments, the transceiver may include a receiver and / or a transmitter, which may be separate or integrated. Optionally, the terms transceiver, transceiver unit, transceiver, transceiver circuit, interface circuit, interface, etc., can be used interchangeably; the terms transmitter, transmitting unit, transmitter, transmitting circuit, etc., can be used interchangeably; and the terms receiver, receiving unit, receiver, receiving circuit, etc., can be used interchangeably.

[0375] In some embodiments, the communication device 6100 further includes one or more memories 6103 for storing data. Optionally, all or part of the memories 6103 may be located outside the communication device 6100. In optional embodiments, the communication device 6100 may include one or more interface circuits 6104. Optionally, the interface circuits 6104 are connected to the memories 6102 and can be used to receive data from the memories 6102 or other devices, and to send data to the memories 6102 or other devices. For example, the interface circuits 6104 can read data stored in the memories 6102 and send the data to the processor 6101.

[0376] The communication device 6100 described in the above embodiments may be a network device or a terminal, but the scope of the communication device 6100 described in this disclosure is not limited thereto, and the structure of the communication device 6100 may not be limited by FIG. 6A. The communication device may be a standalone device or a part of a larger device. For example, the communication device may be: (1) a standalone integrated circuit IC, or chip, or chip system or subsystem; (2) a collection of one or more ICs, optionally, the IC collection may also include storage components for storing data and programs; (3) an ASIC, such as a modem; (4) a module that can be embedded in other devices; (5) a receiver, terminal device, smart terminal device, cellular phone, wireless device, handheld device, mobile unit, vehicle device, network device, cloud device, artificial intelligence device, etc.; (6) others, etc.

[0377] Figure 6B is a schematic diagram of the structure of chip 6200 according to an embodiment of this disclosure. For cases where the communication device 6100 can be a chip or a chip system, please refer to the schematic diagram of chip 6200 shown in Figure 6B, but it is not limited thereto.

[0378] Chip 6200 includes one or more processors 6201. Chip 6200 is used to perform any of the methods described above.

[0379] In some embodiments, chip 6200 further includes one or more interface circuits 6202. Optionally, terms such as interface circuit, interface, and transceiver pin can be used interchangeably. In some embodiments, chip 6200 further includes one or more memories 6203 for storing data. Optionally, all or part of the memories 6203 may be located outside chip 6200. Optionally, interface circuit 6202 is connected to memory 6203, and interface circuit 6202 can be used to receive data from memory 6203 or other devices, and interface circuit 6202 can be used to send data to memory 6203 or other devices. For example, interface circuit 6202 can read data stored in memory 6203 and send the data to processor 6201.

[0380] In some embodiments, the interface circuit 6202 performs at least one of the communication steps, such as sending and / or receiving, in the above-described method. For example, the interface circuit 6202 performing the communication steps, such as sending and / or receiving, in the above-described method means that the interface circuit 6202 performs data interaction between the processor 6201, the chip 6200, the memory 6203, or the transceiver device. In some embodiments, the processor 6201 performs at least one of the other steps.

[0381] The modules and / or devices described in the various embodiments, such as virtual devices, physical devices, and chips, can be combined or separated arbitrarily as needed. Optionally, some or all steps can also be performed collaboratively by multiple modules and / or devices, which is not limited here.

[0382] This disclosure also proposes a storage medium storing instructions that, when executed on the communication device 6100, cause the communication device 6100 to perform any of the above methods. Optionally, the storage medium is an electronic storage medium. Optionally, the storage medium is a computer-readable storage medium, but not limited thereto; it may also be a storage medium readable by other devices. Optionally, the storage medium may be a non-transitory storage medium, but not limited thereto; it may also be a temporary storage medium.

[0383] This disclosure also provides a program product that, when executed by the communication device 6100, causes the communication device 6100 to perform any of the above methods. Optionally, the program product is a computer program product.

[0384] This disclosure also proposes a computer program that, when run on a computer, causes the computer to perform any of the above methods.

Claims

1. A communication method, characterized in that, The method, executed by a terminal, includes: Send first information to the network device, the first information being used to indicate a first model structure, the first model structure being a model structure for which the terminal needs the network device to transmit model parameters; The system receives second information sent by a network device, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

2. The method according to claim 1, characterized in that, The model parameters of the first model include only the model parameters corresponding to the first sub-model structure. The first sub-model structure includes at least one of the following: a portion of the neural network layers in the first model, a portion of the neurons in the first model, a portion of the channels in the first model, and a portion of the convolutional kernels in the first model.

3. The method according to claim 1 or 2, characterized in that, The model parameters of the first model are pruned model parameters.

4. The method according to claim 3, characterized in that, The second information is also used to indicate the pruning content; The pruning includes at least one of the following: The model parameters of the first model are subject to weight pruning. The pruning includes a first neuron and / or a first neural network layer. The first neuron is the neuron whose model parameters are pruned in the first model, and the first neural network layer is the neural network layer whose model parameters are pruned in the first model. The model parameters of the first model are pruned by neurons. The pruning includes the first neuron, which is the neuron in the first model whose corresponding model parameters are pruned. The model parameters of the first model are pruned by convolution kernels. The pruning includes the first convolution kernel, which is the convolution kernel in the first model whose corresponding model parameters are pruned. The model parameters of the first model are structurally pruned. The pruning includes a first neural network layer and / or a second model parameter. The first neural network layer is the neural network layer in the first model whose corresponding model parameters are pruned. The second model parameter is the model parameter between the neural network layers in the first model that is pruned. The model parameters of the first model are pruned through channels. The pruning includes the first channel, which is the channel in the first model where the corresponding model parameters are pruned.

5. The method according to any one of claims 1-4, characterized in that, The model parameters of the first model are the quantized model parameters; In this model, the neural network layer or neurons are divided into multiple parts, and the model parameters corresponding to each part correspond to different quantization precisions.

6. The method according to any one of claims 1-5, characterized in that, The first model includes multiple models with different model structures. The multiple models include a second sub-model structure with the same model structure. The model parameters corresponding to the second sub-model structure are the same. The second sub-model structure includes at least one of the following: neural network layer; neuron; channel; convolution kernel.

7. The method according to any one of claims 1-6, characterized in that, The transmission rank is greater than 1, and the model parameters of the first model satisfy at least one of the following: The model parameters of the models corresponding to multiple or all data transmission layers are the same; The models corresponding to multiple or all data transmission layers include a second sub-model structure with the same model structure and the same model parameters for the second sub-model structure. The second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

8. The method according to any one of claims 1-7, characterized in that, The first information includes an identifier corresponding to each of the first model structures, and the identifier is any one of the following: model identifier, encoder identifier, decoder identifier, or pairing identifier.

9. The method according to any one of claims 1-8, characterized in that, The model parameters of the first model include at least one model parameter corresponding to the structure of the first model, and the second information is also used to indicate the model structure corresponding to each of the at least one model parameter.

10. The method according to any one of claims 1-9, characterized in that, The second information is also used to indicate a second model structure, which is a model structure for which the second information does not indicate the corresponding model parameters in the first model structure.

11. A communication method, characterized in that, Performed by a network device, the method includes: The receiving terminal sends first information, which is used to indicate a first model structure, and the first model structure is a model structure for which the terminal needs the network device to transmit model parameters. Send a second message to the terminal, the second message being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

12. The method according to claim 11, characterized in that, The model parameters of the first model include only the model parameters corresponding to the first sub-model structure. The first sub-model structure includes at least one of the following: a portion of the neural network layers in the first model, a portion of the neurons in the first model, a portion of the channels in the first model, and a portion of the convolutional kernels in the first model.

13. The method according to claim 11 or 12, characterized in that, The model parameters of the first model are pruned model parameters.

14. The method according to claim 13, characterized in that, The second information is also used to indicate the pruning content; The pruning includes at least one of the following: The model parameters of the first model are subject to weight pruning. The pruning includes a first neuron and / or a first neural network layer. The first neuron is the neuron whose model parameters are pruned in the first model, and the first neural network layer is the neural network layer whose model parameters are pruned in the first model. The model parameters of the first model are pruned by neurons. The pruning includes the first neuron, which is the neuron in the first model whose corresponding model parameters are pruned. The model parameters of the first model are pruned by convolution kernels. The pruning includes the first convolution kernel, which is the convolution kernel in the first model whose corresponding model parameters are pruned. The model parameters of the first model are structurally pruned. The pruning includes a first neural network layer and / or a second model parameter. The first neural network layer is the neural network layer in the first model whose corresponding model parameters are pruned. The second model parameter is the model parameter between the neural network layers in the first model that is pruned. The model parameters of the first model are pruned through channels. The pruning includes the first channel, which is the channel in the first model where the corresponding model parameters are pruned.

15. The method according to any one of claims 12-14, characterized in that, The model parameters of the first model are the quantized model parameters; In this model, the neural network layer or neurons are divided into multiple parts, and the model parameters corresponding to each part correspond to different quantization precisions.

16. The method according to any one of claims 12-15, characterized in that, The first model includes multiple models with different model structures. The multiple models include a second sub-model structure with the same model structure. The model parameters corresponding to the second sub-model structure are the same. The second sub-model structure includes at least one of the following: neural network layer; neuron; channel; convolution kernel.

17. The method according to any one of claims 11-16, characterized in that, The transmission rank is greater than 1, and the model parameters of the first model satisfy at least one of the following: The model parameters of the models corresponding to multiple or all data transmission layers are the same; The models corresponding to multiple or all data transmission layers include a second sub-model structure with the same model structure and the same model parameters for the second sub-model structure. The second sub-model structure includes at least one of the following: a neural network layer; a neuron; a channel; and a convolutional kernel.

18. The method according to any one of claims 11-17, characterized in that, The first information includes an identifier corresponding to each of the first model structures, and the identifier is any one of the following: model identifier, encoder identifier, decoder identifier, or pairing identifier.

19. The method according to any one of claims 11-18, characterized in that, The model parameters of the first model include at least one model parameter corresponding to the structure of the first model, and the second information is also used to indicate the model structure corresponding to each of the at least one model parameter.

20. The method according to any one of claims 11-19, characterized in that, The second information is also used to indicate a second model structure, which is a model structure for which the second information does not indicate the corresponding model parameters in the first model structure.

21. A communication device, characterized in that, include: The transceiver module is used to send first information to the network device. The first information is used to indicate a first model structure. The first model structure is a model structure for which the terminal needs the network device to transmit model parameters. The transceiver module is further configured to receive second information sent by the network device, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

22. A communication device, characterized in that, include: The transceiver module is used to receive first information sent by the terminal. The first information is used to indicate a first model structure, which is a model structure for which the terminal needs the network device to transmit model parameters. The transceiver module is further configured to send second information to the terminal, the second information being used to indicate the model parameters of the first model, the model structure of the first model being the first model structure.

23. A communication device, characterized in that, include: One or more processors; The communication device is used to perform the communication method according to any one of claims 1-10 or any one of claims 11-20.

24. A storage medium storing instructions, characterized in that, When the instruction is executed on the communication device, the communication device performs the communication method as claimed in any one of claims 1-10 or any one of claims 11-20.

25. A computer program product comprising a computer program and / or instructions, characterized in that, When the computer program and / or the instructions are executed by the communication device, they implement the communication method as claimed in any one of claims 1-10 or any one of claims 11-20.

Images